Reversible Column Networks
Authors: Yuxuan Cai, Yizhuang Zhou, Qi Han, Jianjian Sun, Xiangwen Kong, Jun Li, Xiangyu Zhang
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments suggest that CNN-style Rev Col models can achieve very competitive performances on multiple computer vision tasks such as image classification, object detection and semantic segmentation, especially with large parameter budget and large dataset. For example, after Image Net-22K pre-training, Rev Col XL obtains 88.2% Image Net-1K accuracy. Given more pre-training data, our largest model Rev Col-H reaches 90.0% on Image Net-1K, 63.8% APbox on COCO detection minival set, 61.0% m Io U on ADE20k segmentation. |
| Researcher Affiliation | Collaboration | Yuxuan Cai1 Yizhuang Zhou1 Qi Han1 Jianjian Sun1 Xiangwen Kong1 Jun Li1 Xiangyu Zhang12 MEGVII Technology1 Beijing Academy of Artificial Intelligence2 |
| Pseudocode | No | No pseudocode or algorithm blocks are explicitly labeled or presented. |
| Open Source Code | Yes | We release code and models at https://github.com/megvii-research/RevCol |
| Open Datasets | Yes | We conduct image classification on Image Net dataset (Deng et al., 2009; Ridnik et al., 2021). We also test our models on the downstream object detection task and semantic segmentation task on commonly used MS-COCO (Lin et al., 2014) and ADE20k (Zhou et al., 2017b) dataset. |
| Dataset Splits | Yes | In Tab. 1, we compare our Rev Col variants with commonly used recent Transformers and CNNs on Image Net-1k validation set. ... We evaluate our proposed Rev Col on object detection task. Experiments are conducted on the MS-COCO dataset using the Cascade Mask R-CNN (Cai & Vasconcelos, 2019) framework. ... APbox on COCO detection minival set. |
| Hardware Specification | Yes | Our experiments are conducted on Nvidia Tesla V100 GPU under batch-size 64, FP16 precision and Py Torch implementation. |
| Software Dependencies | No | The paper mentions "Py Torch implementation" but does not specify a version number or other software dependencies with their versions. |
| Experiment Setup | Yes | F.2 HYPERPARAMETERS USED FOR TRAINING AND PRE-TRAINING This section introduces the training details for main experiments, the supervised training on Image Net and extra data. We show this setting in Tab. 11. All experiments in ablation studies are superivised trained on Image Net-1K except additional descriptions and also follow settings described in this section. (Tables 11, 12, 13, 14 provide specific hyperparameters). |